Evaluating Urban Partitioning Approaches to Improve Crime Forecasting Accuracy in Cities

Urban criminal activity affects public safety, economic stability, and residents' quality of life, making the efficient allocation of public security resources essential. Nowadays, thanks to the ICT penetration and government initiatives, many city administrations are collecting and making available to the community historical data about many aspects of the cities, such as crime events, which can be analyzed with machine learning to create predictive models for crime forecasting. These models can be effectively exploited to optimize resource utilization and aid in targeted crime prevention. Crime rates and patterns vary by geography and time, requiring the partitioning of city territories based on crime trends. While traditional methods use static-defined police districts, machine learning algorithms (especially multi-density clustering methods) can identify crime hotspots more effectively. This paper evaluates the effectiveness of traditional static police district partitioning versus multi-density clustering algorithms (HDBSCAN, CHD) for detecting crime hotspots. Using Chicago crime data as a case study, it then applies the state-of-the-art SARIMA model to forecast crime trends within each identified hotspot. The experimental results highlight the significant advantages of using multi-density clustering over static police district boundaries, offering improved spatial insights and more accurate crime predictions.